Responsible AI Due Diligence Tool
The AI application and infrastructure stack is evolving at exceptional speed. This framework is a v1 due-diligence tool developed with international input from GPs, LPs, operators, and domain experts. It is intentionally agile and iterative: future updates will refine the questions as technologies mature, new risks and opportunities emerge, and our collective understanding of AI’s impact on markets and societies deepens.
This is an evolving project. We welcome your suggestions on both content and format.
Developed by Reframe Venture and ImpactVC with our partner Project Liberty Institute, with support from Zendesk.
Last updated: December 2025
Download as .xlsx
Tell us about a company doing this well
Good Governance for AI Companies
All stages:
What existing and emerging AI regulations or compliance requirements are you aware of that could affect your business?
How are you preparing for regulatory changes, and what would compliance look like at your current stage versus when you scale?
All stages:
Who in your organisation owns AI safety decisions?
How do they interact with product, legal, and business teams?
Pre-Seed / Seed:
Do you have a responsible AI policy in place? What does it cover?
What steps are you taking to make your data and model processes transparent and well-documented?
Do you have any form of model cards, data sheets, or documentation that describes what your AI, how it was trained, and its limitations?
Series A+:
Show us your data and model documentation system (model cards, data cards).
How do you maintain audit trails for AI decisions that would support compliance or incident investigation?
Data Input and Technical Foundation
All stages:
Walk us through your data sources.
What do you own vs. license vs. scrape?
What happens if access to these sources gets restricted or costs increase significantly?
How representative are these sources of your target users?
All stages:
How do you handle sensitive data in your AI systems?
What are your data retention, deletion, and access control policies?
What's your plan if personal data gets leaked through model outputs or attacks?
If an agentic system, how does the system manage and act with personal or sensitive data?
Pre-Seed / Seed:
What third-party AI services/models are you using?
How do you evaluate their reliability at the integration stage, and how do you plan to monitor change in terms or performance?
How easily can you switch between AI models? Can users choose their model?
Series A+:
How do you evaluate and monitor the responsible AI practices of your AI suppliers and partners?
What happens if a vendor fails your standards?
How easily can you switch between AI models? Can users choose their model?
Pre-Seed / Seed:
How are you thinking about strategically balancing model quality with cost?
How do you think about the trade-offs between model performance and efficiency?
What's your current approach to measuring or estimating your AI system's energy and water consumption?
Series A+:
How does energy efficiency factor into your model selection and infrastructure decision; is GPU access a concern?
Show us your water, energy, and carbon monitoring infrastructure. What tools do you use and what metrics do you track?
How do you benchmark your models' energy performance against alternatives, and how does this inform procurement or B2B sales conversations?
Output, Performance, and Algorithm
All stages:
What happens when your AI fails or produces incorrect information?
If agentic, how does your system manage cascading failures? How will you navigate liability?
How do users experience failures, and what safeguards prevent them from acting on wrong AI outputs?
All stages:
What are some sensitive bottlenecks in which your AI driven decision can harm users if biased?
What user groups may be effected by bias in your AI?
How did you test your AI for bias?
All stages:
Walk us through your testing process before deploying AI changes.
How do you validate that your AI system works correctly, safely and without bias before real users interact with it?
What scenarios do you test for?
Do you have sandbox (test) environments or limited user testing to catch issues between development and full deployment?
All stages:
How do you monitor your AI system's performance and behavior after it's live with users?
What happens when you discover the AI has made errors or caused problems, or acted in a bias manner, and how do you respond?
How do feedback loops from monitoring influence your development process?
Designing for adoption
All stages:
How do you communicate AI capabilities and limitations to users?
How are you thinking about building trust amongst customers?
Pre-Seed / Seed:
Where exactly does human judgment end and AI decision-making begin in your product?
What decisions should never be fully automated, or what mitigation procesess do you have to ensure any such delegation of decision making to AI is done properly?
Series A+:
Show us how you've systematically implemented human-in-the-loop design across your product, or explain how you were able to avoid its necessity in risky automation decision points.
What governance processes ensure human oversight boundaries are maintained as you scale?
How do you train and support users in effective human-AI collaboration?
All stages:
How do you see the human AI interaction designed in your product?
Is your AI integration supportive for human bias, machine bias, and biases that can stem from human and AI interactions?
Pre-Seed / Seed:
Who could be harmed if your AI makes systematic errors or is misused by bad actors?
What unintended use cases could emerge that you haven't designed for?
How would you detect these issues happening?
Series A+:
What monitoring systems do you have to detect malicious use patterns, and how do you respond when harmful use cases emerge?
Pre-Seed / Seed:
What records do you keep about your AI decisions?
If a customer asks 'why did the AI do X,' can you explain it?
In what bottlenecks in your product design such decisions' explanations are crucial?
Series A+:
Demonstrate your systematic approach to AI explainability across different user types and use cases.
How do you ensure explanations remain meaningful and actionable as your models become more complex?
What infrastructure supports explainability at scale?
